[lxc-users] OOMScoreAdjust= in dbus.service on systemd-based unprivileged containers

[Christian Brauner]

Hello,

In the past I had troubles getting systemd-based unprivileged containers
such as Archlinux or Debian Jessie to boot. On the other hand some
systemd-based unpriv. containers such as Ubuntu Vivid worked fine and
out of the box. (But I remember Stephane once saying that they were
somehow patched.)

With a little trick I can now also get Archlinux unpriv. containers
booting out of the box, meaning I pacstrap a file-system and then use
uidmapshift to map ids and gids or create a privileged one with the LXC
template and then uidmapshift them.

For this to work the dbus.service file in the unpriv.  Archlinux
container under /usr/lib/systemd/system/ must be patched. It contains
the entry:

    OOMScoreAdjust=-900

where OOMScoreAdjust

    "Sets the adjustment level for the Out-Of-Memory killer for executed
    processes. Takes an integer between -1000 (to disable OOM killing
    for this process) and 1000 (to make killing of this process under
    memory pressure very likely). See proc.txt for details."

(I won't get into detail here as all the relevant parts can be found
under: https://www.kernel.org/doc/Documentation/filesystems/proc.txt)

The OOMScoreAdjust score of each process can be seen by doing

    cat /proc/PID/oom_adj
    cat /proc/PID/oom_score
    cat /proc/PID/oom_score_adj

E.g. the values for dbus-daemon on an Archlinux host will look like
this:

    cat /proc/PID/oom_adj

will be -15 (It is only there for backwards compatibility with older
kernels where the oom adjustement was made there.).

    cat /proc/PID/oom_adj_score

will be -900 as specified in dbus.service on Archlinux.

Finally,

    cat /proc/PID/oom_score

will be 0.

Both values, -15 and -900 make dbus one of the few processes that is
killed last when out of memory.

All unprivileged containers I have run so far (Ubuntu Vivid, Debian
Jessie) even though they may have the line OOMScoreAdjust in their
dbus.service file do not adjust this value meaning when they boot they
show 0 on all three fields under proc.

When disabling OOMScoreAdjust in the Archlinux dbus.service file of the
unpriv. container the boot will be immediately succesfull and all three
values under proc will show 0 just as for an Ubuntu Vivid or Debian
Jessie unpriv. container.

So it seems that during boot systemd in Archlinux is trying to adjust
the oom score under its /proc tree for dbus but somehow fails.

Maybe this will help some people. Should I also file this on LXC github?

Best,
Christian
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.linuxcontainers.org/pipermail/lxc-users/attachments/20150204/80d363ea/attachment.sig>